DocuSign, NVIDIA Tackle Unstructured Agreement Data

Enterprises are facing a significant challenge in managing massive volumes of unstructured agreement data. Crucial business information, such as pricing tables, often remains inaccessible within these documents, which are stored in diverse, unstructured formats.

Hiral Shah stated, "So today Sean and I are going to talk about a massive problem that every enterprise faces which is agreement data largecale agreement data."

DocuSign handles an immense scale of operations, serving 1.9 million paying customers and one billion users, processing approximately one million agreements daily. Despite this volume, a Deloitte study highlights that about $2 trillion in negotiated value remains uncaptured due to manual workflows and disconnected systems.

Hiral Shah emphasized, "we have 1.9 million customers who are paying us and a billion users what does that imply we process us a million agreements a day that needs to now structurize make it readable make it querable usable"

The platform covers agreement creation, negotiation, redlining, and storage, using Generative AI to transform agreements into queryable insights.

Hiral Shah explained, "So that's kind of why docysine built AM an intelligent agreement management platform that takes the entire like applying in an AI first way the entire agreement life cycle"

Agreements are inherently hierarchical and not flat, containing nested dependencies within documents like PDFs and PNGs. Traditional document extraction tools and generic large language models (LLMs) struggle significantly with these complex structures, particularly tables with merged cells.

Their line-by-line reading approach breaks the logical structure of the data, making accurate extraction challenging. Hiral Shah noted, "traditional document extraction tools or a generic VM BLM completely fail here like you know we've tried we've definitely done this because they're reading text line by line which breaks a lot of that concept within the table."

This collaboration focuses on enhancing accuracy by deeply understanding document layouts and reducing the operational overhead for critical teams like legal and procurement.

Hiral Shah stated, "that's kind of where we partnered with Nvidia and leveraged a purpose-built model like you know tool which is literally like you know making our architecture for table extraction take us to make and solve these complex use cases."

The project provides blueprints, datasets, and optimization techniques such as quantization and pruning.

It employs a two-stage approach involving initial document retrieval followed by fine-grained extraction, designed to efficiently handle petabyte-scale enterprise data. Sean Sodha outlined, "so Neimatron is all about building worldclass open-source models and publishing the data sets, the techniques, uh the quantization approaches, distillation approaches, pruning approaches, every technique possible, blueprints to go with that"

Table extraction is a highly complex technique due to the vast number of permutations and structures tables can take, including nested tables and merged cells. These complexities require specialized methods that go beyond basic extraction techniques to accurately parse and understand tabular data.

Sean Sodha highlighted, "table extraction which is a really complicated technique um if you think about it the number of permutations of tables are quite vast"

NVIDIA introduced NeMo Parse, an 850-900 million parameter Vision-Language Model (VLM) specifically designed for document tasks. Unlike typical multi-billion parameter VLMs, NeMo Parse is a smaller, more efficient all-in-one extractor.

It avoids the heavy overhead often associated with larger VLMs while effectively outputting semantic layout, extracted text, and correct reading order from documents. Sean Sodha explained, "when you think VLM you generally think a multi-billion parameter model... this is a very small tiny C radio VLM. It's about 850 900 million parameter model."

Models are benchmarked against standards like RD Table Bench, and NVIDIA is committed to open-sourcing these models and their weights.

Sean Sodha stated, "we want to make sure that both on the paro curve of accuracy versus performance, we're make sure that we're going to be releasing a world-class models to the ecosystem"

A demonstration showcased the Agreement Manager processing an order form, instantly uploading and structuring metadata using AI. Users can navigate directly to specific sections of the agreement, such as key terms and commercial details, which are automatically extracted and organized.

This structured data can then be downloaded, providing clean, actionable information for finance and procurement teams, streamlining their operations. Hiral Shah described the process, "the moment I select the agreement, it starts uploading and starts processing with AI... agreement manager has extracted a rich set of metadata. Everything from key terms to commercial details are automatically structured"

The system scales to handle thousands, even millions, of agreements, allowing specific business units or leaders to query for targeted information.

Proprietary data models provide comprehensive organizational oversight, transforming previously inaccessible data into actionable intelligence. Hiral Shah highlighted, "the big piece is how do we understand and get that data that makes it very valuable to an end business user right a legal person a procurement person a salesperson"

A key learning from the DocuSign-NVIDIA partnership is the critical importance of using purpose-built models over generic ones for specific tasks like document extraction. Model efficiency, characterized by lower parameter counts, significantly contributes to faster performance and quicker market deployment.

This strategic focus on tailored AI solutions has enabled accelerated development and integration of advanced capabilities. Hiral Shah emphasized, "a purpose-built model for the job you're trying to do is a big big part of how we've been thinking about and that's kind of where we've been able to accelerate bring things to market much faster."